In this work, we found that it is essential to divide the training data into proper subsets according to the wild and mutated type of SAVs when the model is trained. Moreover, by the feature selection procedure, the critical descriptors could be figured out. The relationship between the mutated residues and the interaction changed could be studied and characterized, primarily by analyzing the micro-environment-based feature set. Although further study is needed to reveal out the mechanism of cancer in most selected features, our results indicate that it is possible to predict cancer-related SAV reliably. Our work will provide a useful, practical tool for cancer research and precision medicine.